Samsung thinks Claude Code can help it boost chip design — but admits the AI still makes some worryingly big mistakes

Samsung System LSI division uses Anthropic's Claude Code to reportedly slice working processes However the tool also apparently downgraded an error message instead of fixing it, rolling back unrelated finished work, and tried to edit code it was not authorized to touch Samsung's engineers currently review everything the tool produces before it is published elsewhere or affects existing chip designs Samsung's System LSI division has reportedly used Anthropic's Claude Code to cut a month-long system-on-chip verification job to about two days.

While Samsung's division has about 6,000 employees, its main rival in mobile application processors, Qualcomm, has roughly 52,000 employees, and the former is apparently not above using AI to bridge the gap as a force multiplier.

A report from South Korean business outlet Chosun Biz (KR) claims System LSI leveraged Anthropic's Claude Code by offering it to its software developers in May 2026, then extended it to semiconductor design and verification work, a move which led to much faster progress on certain projects, while also producing interesting failures on other fronts.

A mix of triumphs and gaffes Samsung's approach, for the most part, makes an excellent case for future AI use, as one verification project expected to take more than a month was finished in about two days, something the company internally tracked as a 15x gain in efficiency.

More interestingly, a second-year engineer with no prior exposure to either the tool or USB communication standards built USB device models for an emulator and adapted an Android driver in a single day, work whichwas normally estimated to take about a month.

Samsung has also had an interesting interaction using Claude Code to verify the internal data connections of a custom system-on-chip.

The job was complex in more ways than one: the customer wanted a new architecture, third-party design IP was involved, the documentation was nonstandard, and the RTL for the DRAM controller had not arrived on schedule, making for a perfect storm for Claude.

Claude built a virtual verification environment using placeholder blocks in place of the missing RTL, along with test scenarios, so that engineers could start catching errors before the real design existed.

On the other side of the equation, Claude Code did, in certain cases, try to operate outside the boundaries of its assignments, often resulting in incorrect output or an error, or, in one case, an attempt to modify RTL circuit code it had no authorization to touch.

It also rolled back unrelated finished work and downgraded an error message instead of fixing it.

The former is particularly egregious because, unlike software updates, once silicon ships, it can rarely be repurposed or 'updated' to bypass design flaws.